B2b lead finder

B2B Lead Finder

Turn a product + location + evidence keywords into a verified, prioritized lead list. The core promise of this skill is evidence-based leads: every row in the final table exists because a real public source was visited and quoted, not because a name sounded plausible.

Inputs to collect before starting

Ask for whatever is missing; propose sensible defaults for the rest:

1. Product / value proposition — what is being sold, and which capabilities matter (e.g. "secure video meetings, class hosting, recording, webinar hosting").
2. Location — city or region, used as a hard filter.
3. Lead categories — the business types to target (coaching centers, clinics, consultancies…).
4. Mention terms — the phrases a qualifying business would use publicly ("online class", "teleconsultation", "webinar"…). Include local-language equivalents — in many markets the Facebook page is in the local language even when the website is in English.
5. Target count — how many verified leads to aim for.
6. Extra rules — exclusions, priority preferences, required fields.

A library of ready-made vertical × city campaign presets (e.g. IELTS centers in Karachi, dental clinics in Riyadh, SaaS agencies in Berlin) is indexed in references/presets-index.md — check the index FIRST whenever the request matches a known vertical or market; the index is the authoritative list. For uncovered markets, copy references/preset-template.md and fill it from the nearest neighbor.

Founder quick start

Founders usually arrive with a product and a city, not a marketing brief. Don't interrogate them with six questions — derive the brief for them:

1. From their product description (or their website, if they give a URL), propose 3–5 lead categories and 8–12 mention terms that would mark a business as needing this product. Show the proposed config in one compact block and get a yes/edit before spending anything on search.
2. Default the target to 25–50 verified leads, not hundreds. Founders do personal outreach; 40 leads with strong evidence beat 500 unsorted rows they'll never work through.
3. Write why_good_fit as the sentence the founder could open their first email with ("Saw you run nightly IELTS batches over Zoom…") — specific enough to reuse verbatim.
4. Offer to run one category first as a probe. If verification confirms that category is rich in High-priority leads, expand; if it comes back thin, the founder just learned something about their market for the price of one search round.

Non-negotiable data rules

These exist because a lead list is only as valuable as its worst row — one invented phone number or dead business discovered by a salesperson poisons trust in the entire table.

• Never fabricate. If a phone, email, WhatsApp, or LinkedIn page cannot be found on a real public page, leave the cell blank. A blank cell is useful information; a guessed one is a trap.
• Public business data only. Collect contact details a business itself publishes (website contact page, Google Maps listing, Facebook About). Do not collect personal private data unless the business lists it as its official contact.
• Evidence or it didn't happen. Every lead carries the exact quoted sentence where the mention term appears and the URL where it was seen.
• Honest counts. If the target is 500 and only 137 real verified leads exist, deliver 137 and say so. Never pad with unverified rows to hit a number.

Pipeline

Run four phases. On long runs, checkpoint results to disk after every batch — provider rate limits and session caps can kill a run mid-flight, and a checkpointed run resumes for free.

Phase 1 — Scope

Confirm inputs, then create the job folder — ./leads/<vertical>-<city>-<YYYY-MM-DD>/ containing config.md, state.json, leads.csv, and rejects.csv — and tell the user that path in the first status message. Everything the run produces lives there.

Set expectations with real numbers before spending: plan on roughly 5–10 searches plus 3–6 page fetches per verified lead; verification will reject 30–60% of raw candidates (normal and healthy — so discover 2–3× the target count); a 10-lead run takes ~30–60 minutes with plain web search, less with accelerators. Start with 3–5 queries per category per round, and default to a 3-round cap unless the user raises it.

Phase 2 — Discover (recall-oriented)

For each category, search broadly for candidate businesses:

• Combine category keywords × location × mention terms into many query variations.
• Search the open web AND platform-scoped queries (site:facebook.com, site:linkedin.com/company, Google Maps queries). Different platforms surface different businesses — a coaching center may have no website but a very active Facebook page.
• Use local-language mention terms as first-class queries, not an afterthought.
• For each candidate, open its website/Facebook page and record: name, all contact fields found, physical address + neighborhood, the exact mention quote, and the source URL.
• WhatsApp numbers usually hide as wa.me/<number> or api.whatsapp.com links on Facebook pages and website footers — grep fetched pages for those patterns rather than expecting a labeled field.
• Deduplicate across the whole run by normalized name — lowercase, strip punctuation, drop legal suffixes (ltd, limited, inc, llc, & co) — and also by website domain: two spellings of one academy share a domain long before they share a spelling.

Query recipes that work (vary and combine):

"<category keyword>" <city> "<mention term>"
site:facebook.com "<mention term (local language)>" <city>
site:linkedin.com/company <category keyword> <city>
"<mention term>" <city> contact OR admission OR appointment

Discovery optimizes for recall — it will surface plausible-but-wrong candidates. That is fine. Do not polish here; verification is the filter.

Phase 3 — Verify (adversarial, one pass per lead)

Treat every discovery claim as unproven. For each candidate, independently re-visit sources and confirm:

1. Real and active — the business exists and shows recent activity (recent social posts, updated Maps profile, working website). Default thresholds: activity within 6 months = active; 6–18 months = accept but mark stale in notes; older than 18 months with no other signal = reject as inactive. Note the signal either way.
2. In-location — physically inside the target location, with the specific area named. Cannot confirm → reject.
3. Genuine mention — the quoted online-service claim is actually findable at the recorded source (or elsewhere on the business's own pages). Cannot find it → reject.
4. Contact cross-check — at least one of phone / website / Facebook confirmed by a second independent source (Maps listing vs website footer, directory listing, etc.). Note mismatches. Record confidence = how many independent sources confirmed the business's identity and contacts: 1 = single-source (allowed, but say so), 2 = cross-checked, 3+ = solid. Binary verified/rejected hides exactly the information a salesperson uses to decide who to call first.
5. Fill gaps — add any missing public contact fields discovered during verification.

Rejected candidates are kept in a separate rejects file with the rejection reason — the user may loosen criteria later, and rejects prove the filter is real.

Phase 4 — Deliver

Produce leads.csv and rejects.csv plus a short summary: counts by category and priority, rejection rate and top rejection reasons, and honest notes on coverage gaps.

CSV craft that prevents real-world breakage:

• Write UTF-8 with BOM (utf-8-sig in Python). Without the BOM, Excel renders non-Latin text (Bangla, Arabic, Thai…) as mojibake — and local-language quotes are half the value.
• Quote every field; keep phone numbers as text so leading +/0 survive Excel.
• One row per organization, columns exactly as specified below.

Priority rubric

When a preset is loaded, its vertical-specific rubric replaces this generic one — the generic rubric below is the fallback for preset-less runs.

• High — runs frequent/recurring online sessions at meaningful scale (daily classes, a consultation booking system, scheduled webinars) with a real client/student base. The product would replace or upgrade something they demonstrably do every week.
• Medium — mentions online services but appears occasional or small-scale.
• Low — a single or passing mention; unclear ongoing need.

Tie why_good_fit to a specific observed behavior + a specific product capability ("runs live IELTS batches on Zoom nightly → needs stable class hosting + recording"), never a generic line ("could benefit from video conferencing").

Output columns

Use exactly these columns, in this order, unless the user overrides:

organization_name, category, website, google_maps_link, facebook_page, instagram_page, linkedin_page, phone, whatsapp, email, physical_address, area, online_service_mentioned, exact_source_quote, source_url, why_good_fit, priority, confidence, notes

Presets may add market-specific contact columns after whatsapp (e.g. zalo for Vietnam, line for Thailand/Japan) — in many markets the dominant channel isn't in the base set, and a lead list missing the channel businesses actually answer on is much less actionable. Keep the base column order unchanged.

Write rows with scripts/lead_csv.py — append_leads() (or the CLI) enforces column order, UTF-8 BOM, and quote-all in one place, and dedupe_key() / load_state() / save_state() give every run identical dedupe and checkpoint behavior instead of re-implementing them from prose.

Quality bar — one illustrative row (placeholder digits; shape and specificity are the point):

organization_name: Example IELTS Academy
category: IELTS, language, and admission coaching centers
website: https://exampleielts.com.bd
phone: "+8801XXXXXXXXX"          # found on website footer, matched Google Maps listing
area: Dhanmondi
online_service_mentioned: Live online IELTS batches
exact_source_quote: "Join our live online IELTS batch — classes every evening via Zoom."
source_url: https://facebook.com/exampleielts/about
why_good_fit: Runs nightly live IELTS batches on Zoom for ~200 students/month — needs
  reliable class hosting, recording for absentees, and admission-webinar support.
priority: High
confidence: 2          # phone matched on website footer + Google Maps listing
notes: Facebook active (posted 2 days ago); phone verified across website + Maps.

Weak versions of why_good_fit ("could benefit from video conferencing") mean the verification pass didn't look hard enough — send it back.

After delivery — outreach handoff

A lead list is a pipeline stage, not an end product. After delivering, check what outreach tooling exists in the environment and offer the next step instead of stopping:

• Email verification — if an email-verification key or skill is available (MillionVerifier, ZeroBounce…), offer to validate the email column before anyone sends to it.
• List scoring / sequence copy / campaign upload — if cold-outreach skills are installed (list scoring, campaign copywriting, Smartlead/Instantly upload), offer the chain: score the list → draft sequences → upload. why_good_fit is written to serve as each lead's personalization opener — say so when handing off.
• No tooling installed — tell the user the CSV maps directly onto any CRM or outreach platform import (HubSpot, Pipedrive, Smartlead, Instantly), with email, organization_name, and why_good_fit as the fields that matter most.

Sort any handoff by priority then confidence descending — the first leads worked should be the ones most likely to answer.

Running at scale

• If a workflow/subagent orchestration tool is available, fan discovery out one agent per category and verification out one agent per lead (or batches of up to 5 — bigger batches save cost but let one hard-to-verify lead starve the others), in parallel. Pass each discovery agent the list of already-found names so rounds don't re-surface duplicates.
• Checkpoint to a state.json in the job folder after every batch: {"round": N, "seen_names": [...], "confirmed": [...], "rejected": [...]} — plus append confirmed rows to leads.csv as they verify, not only at the end. A rate-limited or interrupted run then resumes by reading state and skipping finished work instead of re-spending on it.
• Loop discover → verify rounds until the target is met, candidates run dry (two consecutive rounds with nothing new), or a round cap is hit.
• Expect the pool to be finite: niche B2B categories in one city often exhaust at a few hundred real businesses. Falling short of an ambitious target is a finding, not a failure.
• Optional accelerators, if installed: the firecrawl CLI for search + scraping, an Apify scraper skill for Google Maps / LinkedIn at volume, a Serper-based Maps skill for structured place data. Use them when present; plain web search + page fetches work when not.

API keys and credentials

This skill ships with no API keys, and none may ever be added to it — it is designed to be shared publicly, and a key committed to a skill leaks to everyone who installs it.

When an accelerator would help but has no working credentials:

1. Check for existing auth first (e.g. firecrawl --status, an env file the tool documents).
2. If missing, ask the user running the skill for their own key, telling them where to get it and roughly what it costs. Never reuse a key found in an unrelated project or paste one from documentation.
3. Save what they provide outside the skill folder — an env file in the user's home directory with chmod 600 — so it persists for them but can never be committed with the skill.
4. If the user has no key, continue with plain web search + page fetches. The pipeline degrades in speed, not in integrity.

Never print a key back in conversation, logs, or output files; when confirming a save, show the variable name only.